1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * linux/kernel/acct.c
4 *
5 * BSD Process Accounting for Linux
6 *
7 * Author: Marco van Wieringen <mvw@planets.elm.net>
8 *
9 * Some code based on ideas and code from:
10 * Thomas K. Dyas <tdyas@eden.rutgers.edu>
11 *
12 * This file implements BSD-style process accounting. Whenever any
13 * process exits, an accounting record of type "struct acct" is
14 * written to the file specified with the acct() system call. It is
15 * up to user-level programs to do useful things with the accounting
16 * log. The kernel just provides the raw accounting information.
17 *
18 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V.
19 *
20 * Plugged two leaks. 1) It didn't return acct_file into the free_filps if
21 * the file happened to be read-only. 2) If the accounting was suspended
22 * due to the lack of space it happily allowed to reopen it and completely
23 * lost the old acct_file. 3/10/98, Al Viro.
24 *
25 * Now we silently close acct_file on attempt to reopen. Cleaned sys_acct().
26 * XTerms and EMACS are manifestations of pure evil. 21/10/98, AV.
27 *
28 * Fixed a nasty interaction with sys_umount(). If the accounting
29 * was suspeneded we failed to stop it on umount(). Messy.
30 * Another one: remount to readonly didn't stop accounting.
31 * Question: what should we do if we have CAP_SYS_ADMIN but not
32 * CAP_SYS_PACCT? Current code does the following: umount returns -EBUSY
33 * unless we are messing with the root. In that case we are getting a
34 * real mess with do_remount_sb(). 9/11/98, AV.
35 *
36 * Fixed a bunch of races (and pair of leaks). Probably not the best way,
37 * but this one obviously doesn't introduce deadlocks. Later. BTW, found
38 * one race (and leak) in BSD implementation.
39 * OK, that's better. ANOTHER race and leak in BSD variant. There always
40 * is one more bug... 10/11/98, AV.
41 *
42 * Oh, fsck... Oopsable SMP race in do_process_acct() - we must hold
43 * ->mmap_lock to walk the vma list of current->mm. Nasty, since it leaks
44 * a struct file opened for write. Fixed. 2/6/2000, AV.
45 */
46
47 #include <linux/mm.h>
48 #include <linux/slab.h>
49 #include <linux/acct.h>
50 #include <linux/capability.h>
51 #include <linux/file.h>
52 #include <linux/tty.h>
53 #include <linux/security.h>
54 #include <linux/vfs.h>
55 #include <linux/jiffies.h>
56 #include <linux/times.h>
57 #include <linux/syscalls.h>
58 #include <linux/mount.h>
59 #include <linux/uaccess.h>
60 #include <linux/sched/cputime.h>
61
62 #include <asm/div64.h>
63 #include <linux/pid_namespace.h>
64 #include <linux/fs_pin.h>
65
66 /*
67 * These constants control the amount of freespace that suspend and
68 * resume the process accounting system, and the time delay between
69 * each check.
70 * Turned into sysctl-controllable parameters. AV, 12/11/98
71 */
72
73 static int acct_parm[3] = {4, 2, 30};
74 #define RESUME (acct_parm[0]) /* >foo% free space - resume */
75 #define SUSPEND (acct_parm[1]) /* <foo% free space - suspend */
76 #define ACCT_TIMEOUT (acct_parm[2]) /* foo second timeout between checks */
77
78 #ifdef CONFIG_SYSCTL
79 static const struct ctl_table kern_acct_table[] = {
80 {
81 .procname = "acct",
82 .data = &acct_parm,
83 .maxlen = 3*sizeof(int),
84 .mode = 0644,
85 .proc_handler = proc_dointvec,
86 },
87 };
88
kernel_acct_sysctls_init(void)89 static __init int kernel_acct_sysctls_init(void)
90 {
91 register_sysctl_init("kernel", kern_acct_table);
92 return 0;
93 }
94 late_initcall(kernel_acct_sysctls_init);
95 #endif /* CONFIG_SYSCTL */
96
97 /*
98 * External references and all of the globals.
99 */
100
101 struct bsd_acct_struct {
102 struct fs_pin pin;
103 atomic_long_t count;
104 struct rcu_head rcu;
105 struct mutex lock;
106 bool active;
107 bool check_space;
108 unsigned long needcheck;
109 struct file *file;
110 struct pid_namespace *ns;
111 struct work_struct work;
112 struct completion done;
113 acct_t ac;
114 };
115
116 static void fill_ac(struct bsd_acct_struct *acct);
117 static void acct_write_process(struct bsd_acct_struct *acct);
118
119 /*
120 * Check the amount of free space and suspend/resume accordingly.
121 */
check_free_space(struct bsd_acct_struct * acct)122 static bool check_free_space(struct bsd_acct_struct *acct)
123 {
124 struct kstatfs sbuf;
125
126 if (!acct->check_space)
127 return acct->active;
128
129 /* May block */
130 if (vfs_statfs(&acct->file->f_path, &sbuf))
131 return acct->active;
132
133 if (acct->active) {
134 u64 suspend = sbuf.f_blocks * SUSPEND;
135 do_div(suspend, 100);
136 if (sbuf.f_bavail <= suspend) {
137 acct->active = false;
138 pr_info("Process accounting paused\n");
139 }
140 } else {
141 u64 resume = sbuf.f_blocks * RESUME;
142 do_div(resume, 100);
143 if (sbuf.f_bavail >= resume) {
144 acct->active = true;
145 pr_info("Process accounting resumed\n");
146 }
147 }
148
149 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
150 return acct->active;
151 }
152
acct_put(struct bsd_acct_struct * p)153 static void acct_put(struct bsd_acct_struct *p)
154 {
155 if (atomic_long_dec_and_test(&p->count))
156 kfree_rcu(p, rcu);
157 }
158
to_acct(struct fs_pin * p)159 static inline struct bsd_acct_struct *to_acct(struct fs_pin *p)
160 {
161 return p ? container_of(p, struct bsd_acct_struct, pin) : NULL;
162 }
163
acct_get(struct pid_namespace * ns)164 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
165 {
166 struct bsd_acct_struct *res;
167 again:
168 smp_rmb();
169 rcu_read_lock();
170 res = to_acct(READ_ONCE(ns->bacct));
171 if (!res) {
172 rcu_read_unlock();
173 return NULL;
174 }
175 if (!atomic_long_inc_not_zero(&res->count)) {
176 rcu_read_unlock();
177 cpu_relax();
178 goto again;
179 }
180 rcu_read_unlock();
181 mutex_lock(&res->lock);
182 if (res != to_acct(READ_ONCE(ns->bacct))) {
183 mutex_unlock(&res->lock);
184 acct_put(res);
185 goto again;
186 }
187 return res;
188 }
189
acct_pin_kill(struct fs_pin * pin)190 static void acct_pin_kill(struct fs_pin *pin)
191 {
192 struct bsd_acct_struct *acct = to_acct(pin);
193 mutex_lock(&acct->lock);
194 /*
195 * Fill the accounting struct with the exiting task's info
196 * before punting to the workqueue.
197 */
198 fill_ac(acct);
199 schedule_work(&acct->work);
200 wait_for_completion(&acct->done);
201 cmpxchg(&acct->ns->bacct, pin, NULL);
202 mutex_unlock(&acct->lock);
203 pin_remove(pin);
204 acct_put(acct);
205 }
206
close_work(struct work_struct * work)207 static void close_work(struct work_struct *work)
208 {
209 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
210 struct file *file = acct->file;
211
212 /* We were fired by acct_pin_kill() which holds acct->lock. */
213 acct_write_process(acct);
214 if (file->f_op->flush)
215 file->f_op->flush(file, NULL);
216 __fput_sync(file);
217 complete(&acct->done);
218 }
219
acct_on(struct filename * pathname)220 static int acct_on(struct filename *pathname)
221 {
222 struct file *file;
223 struct vfsmount *mnt, *internal;
224 struct pid_namespace *ns = task_active_pid_ns(current);
225 struct bsd_acct_struct *acct;
226 struct fs_pin *old;
227 int err;
228
229 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
230 if (!acct)
231 return -ENOMEM;
232
233 /* Difference from BSD - they don't do O_APPEND */
234 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
235 if (IS_ERR(file)) {
236 kfree(acct);
237 return PTR_ERR(file);
238 }
239
240 if (!S_ISREG(file_inode(file)->i_mode)) {
241 kfree(acct);
242 filp_close(file, NULL);
243 return -EACCES;
244 }
245
246 /* Exclude kernel kernel internal filesystems. */
247 if (file_inode(file)->i_sb->s_flags & (SB_NOUSER | SB_KERNMOUNT)) {
248 kfree(acct);
249 filp_close(file, NULL);
250 return -EINVAL;
251 }
252
253 /* Exclude procfs and sysfs. */
254 if (file_inode(file)->i_sb->s_iflags & SB_I_USERNS_VISIBLE) {
255 kfree(acct);
256 filp_close(file, NULL);
257 return -EINVAL;
258 }
259
260 if (!(file->f_mode & FMODE_CAN_WRITE)) {
261 kfree(acct);
262 filp_close(file, NULL);
263 return -EIO;
264 }
265 internal = mnt_clone_internal(&file->f_path);
266 if (IS_ERR(internal)) {
267 kfree(acct);
268 filp_close(file, NULL);
269 return PTR_ERR(internal);
270 }
271 err = mnt_get_write_access(internal);
272 if (err) {
273 mntput(internal);
274 kfree(acct);
275 filp_close(file, NULL);
276 return err;
277 }
278 mnt = file->f_path.mnt;
279 file->f_path.mnt = internal;
280
281 atomic_long_set(&acct->count, 1);
282 init_fs_pin(&acct->pin, acct_pin_kill);
283 acct->file = file;
284 acct->needcheck = jiffies;
285 acct->ns = ns;
286 mutex_init(&acct->lock);
287 INIT_WORK(&acct->work, close_work);
288 init_completion(&acct->done);
289 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
290 pin_insert(&acct->pin, mnt);
291
292 rcu_read_lock();
293 old = xchg(&ns->bacct, &acct->pin);
294 mutex_unlock(&acct->lock);
295 pin_kill(old);
296 mnt_put_write_access(mnt);
297 mntput(mnt);
298 return 0;
299 }
300
301 static DEFINE_MUTEX(acct_on_mutex);
302
303 /**
304 * sys_acct - enable/disable process accounting
305 * @name: file name for accounting records or NULL to shutdown accounting
306 *
307 * sys_acct() is the only system call needed to implement process
308 * accounting. It takes the name of the file where accounting records
309 * should be written. If the filename is NULL, accounting will be
310 * shutdown.
311 *
312 * Returns: 0 for success or negative errno values for failure.
313 */
SYSCALL_DEFINE1(acct,const char __user *,name)314 SYSCALL_DEFINE1(acct, const char __user *, name)
315 {
316 int error = 0;
317
318 if (!capable(CAP_SYS_PACCT))
319 return -EPERM;
320
321 if (name) {
322 struct filename *tmp = getname(name);
323
324 if (IS_ERR(tmp))
325 return PTR_ERR(tmp);
326 mutex_lock(&acct_on_mutex);
327 error = acct_on(tmp);
328 mutex_unlock(&acct_on_mutex);
329 putname(tmp);
330 } else {
331 rcu_read_lock();
332 pin_kill(task_active_pid_ns(current)->bacct);
333 }
334
335 return error;
336 }
337
acct_exit_ns(struct pid_namespace * ns)338 void acct_exit_ns(struct pid_namespace *ns)
339 {
340 rcu_read_lock();
341 pin_kill(ns->bacct);
342 }
343
344 /*
345 * encode an u64 into a comp_t
346 *
347 * This routine has been adopted from the encode_comp_t() function in
348 * the kern_acct.c file of the FreeBSD operating system. The encoding
349 * is a 13-bit fraction with a 3-bit (base 8) exponent.
350 */
351
352 #define MANTSIZE 13 /* 13 bit mantissa. */
353 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
354 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
355
encode_comp_t(u64 value)356 static comp_t encode_comp_t(u64 value)
357 {
358 int exp, rnd;
359
360 exp = rnd = 0;
361 while (value > MAXFRACT) {
362 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
363 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
364 exp++;
365 }
366
367 /*
368 * If we need to round up, do it (and handle overflow correctly).
369 */
370 if (rnd && (++value > MAXFRACT)) {
371 value >>= EXPSIZE;
372 exp++;
373 }
374
375 if (exp > (((comp_t) ~0U) >> MANTSIZE))
376 return (comp_t) ~0U;
377 /*
378 * Clean it up and polish it off.
379 */
380 exp <<= MANTSIZE; /* Shift the exponent into place */
381 exp += value; /* and add on the mantissa. */
382 return exp;
383 }
384
385 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
386 /*
387 * encode an u64 into a comp2_t (24 bits)
388 *
389 * Format: 5 bit base 2 exponent, 20 bits mantissa.
390 * The leading bit of the mantissa is not stored, but implied for
391 * non-zero exponents.
392 * Largest encodable value is 50 bits.
393 */
394
395 #define MANTSIZE2 20 /* 20 bit mantissa. */
396 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
397 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
398 #define MAXEXP2 ((1 << EXPSIZE2) - 1) /* Maximum exponent. */
399
encode_comp2_t(u64 value)400 static comp2_t encode_comp2_t(u64 value)
401 {
402 int exp, rnd;
403
404 exp = (value > (MAXFRACT2>>1));
405 rnd = 0;
406 while (value > MAXFRACT2) {
407 rnd = value & 1;
408 value >>= 1;
409 exp++;
410 }
411
412 /*
413 * If we need to round up, do it (and handle overflow correctly).
414 */
415 if (rnd && (++value > MAXFRACT2)) {
416 value >>= 1;
417 exp++;
418 }
419
420 if (exp > MAXEXP2) {
421 /* Overflow. Return largest representable number instead. */
422 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
423 } else {
424 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
425 }
426 }
427 #elif ACCT_VERSION == 3
428 /*
429 * encode an u64 into a 32 bit IEEE float
430 */
encode_float(u64 value)431 static u32 encode_float(u64 value)
432 {
433 unsigned exp = 190;
434 unsigned u;
435
436 if (value == 0)
437 return 0;
438 while ((s64)value > 0) {
439 value <<= 1;
440 exp--;
441 }
442 u = (u32)(value >> 40) & 0x7fffffu;
443 return u | (exp << 23);
444 }
445 #endif
446
447 /*
448 * Write an accounting entry for an exiting process
449 *
450 * The acct_process() call is the workhorse of the process
451 * accounting system. The struct acct is built here and then written
452 * into the accounting file. This function should only be called from
453 * do_exit() or when switching to a different output file.
454 */
455
fill_ac(struct bsd_acct_struct * acct)456 static void fill_ac(struct bsd_acct_struct *acct)
457 {
458 struct pacct_struct *pacct = ¤t->signal->pacct;
459 struct file *file = acct->file;
460 acct_t *ac = &acct->ac;
461 u64 elapsed, run_time;
462 time64_t btime;
463 struct tty_struct *tty;
464
465 lockdep_assert_held(&acct->lock);
466
467 if (time_is_after_jiffies(acct->needcheck)) {
468 acct->check_space = false;
469
470 /* Don't fill in @ac if nothing will be written. */
471 if (!acct->active)
472 return;
473 } else {
474 acct->check_space = true;
475 }
476
477 /*
478 * Fill the accounting struct with the needed info as recorded
479 * by the different kernel functions.
480 */
481 memset(ac, 0, sizeof(acct_t));
482
483 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
484 strscpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
485
486 /* calculate run_time in nsec*/
487 run_time = ktime_get_ns();
488 run_time -= current->group_leader->start_time;
489 /* convert nsec -> AHZ */
490 elapsed = nsec_to_AHZ(run_time);
491 #if ACCT_VERSION == 3
492 ac->ac_etime = encode_float(elapsed);
493 #else
494 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
495 (unsigned long) elapsed : (unsigned long) -1l);
496 #endif
497 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
498 {
499 /* new enlarged etime field */
500 comp2_t etime = encode_comp2_t(elapsed);
501
502 ac->ac_etime_hi = etime >> 16;
503 ac->ac_etime_lo = (u16) etime;
504 }
505 #endif
506 do_div(elapsed, AHZ);
507 btime = ktime_get_real_seconds() - elapsed;
508 ac->ac_btime = clamp_t(time64_t, btime, 0, U32_MAX);
509 #if ACCT_VERSION == 2
510 ac->ac_ahz = AHZ;
511 #endif
512
513 spin_lock_irq(¤t->sighand->siglock);
514 tty = current->signal->tty; /* Safe as we hold the siglock */
515 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
516 ac->ac_utime = encode_comp_t(nsec_to_AHZ(pacct->ac_utime));
517 ac->ac_stime = encode_comp_t(nsec_to_AHZ(pacct->ac_stime));
518 ac->ac_flag = pacct->ac_flag;
519 ac->ac_mem = encode_comp_t(pacct->ac_mem);
520 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
521 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
522 ac->ac_exitcode = pacct->ac_exitcode;
523 spin_unlock_irq(¤t->sighand->siglock);
524
525 /* we really need to bite the bullet and change layout */
526 ac->ac_uid = from_kuid_munged(file->f_cred->user_ns, current_uid());
527 ac->ac_gid = from_kgid_munged(file->f_cred->user_ns, current_gid());
528 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
529 /* backward-compatible 16 bit fields */
530 ac->ac_uid16 = ac->ac_uid;
531 ac->ac_gid16 = ac->ac_gid;
532 #elif ACCT_VERSION == 3
533 {
534 struct pid_namespace *ns = acct->ns;
535
536 ac->ac_pid = task_tgid_nr_ns(current, ns);
537 rcu_read_lock();
538 ac->ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
539 rcu_read_unlock();
540 }
541 #endif
542 }
543
acct_write_process(struct bsd_acct_struct * acct)544 static void acct_write_process(struct bsd_acct_struct *acct)
545 {
546 struct file *file = acct->file;
547 const struct cred *cred;
548 acct_t *ac = &acct->ac;
549
550 /* Perform file operations on behalf of whoever enabled accounting */
551 cred = override_creds(file->f_cred);
552
553 /*
554 * First check to see if there is enough free_space to continue
555 * the process accounting system. Then get freeze protection. If
556 * the fs is frozen, just skip the write as we could deadlock
557 * the system otherwise.
558 */
559 if (check_free_space(acct) && file_start_write_trylock(file)) {
560 /* it's been opened O_APPEND, so position is irrelevant */
561 loff_t pos = 0;
562 __kernel_write(file, ac, sizeof(acct_t), &pos);
563 file_end_write(file);
564 }
565
566 revert_creds(cred);
567 }
568
do_acct_process(struct bsd_acct_struct * acct)569 static void do_acct_process(struct bsd_acct_struct *acct)
570 {
571 unsigned long flim;
572
573 /* Accounting records are not subject to resource limits. */
574 flim = rlimit(RLIMIT_FSIZE);
575 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
576 fill_ac(acct);
577 acct_write_process(acct);
578 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
579 }
580
581 /**
582 * acct_collect - collect accounting information into pacct_struct
583 * @exitcode: task exit code
584 * @group_dead: not 0, if this thread is the last one in the process.
585 */
acct_collect(long exitcode,int group_dead)586 void acct_collect(long exitcode, int group_dead)
587 {
588 struct pacct_struct *pacct = ¤t->signal->pacct;
589 u64 utime, stime;
590 unsigned long vsize = 0;
591
592 if (group_dead && current->mm) {
593 struct mm_struct *mm = current->mm;
594 VMA_ITERATOR(vmi, mm, 0);
595 struct vm_area_struct *vma;
596
597 mmap_read_lock(mm);
598 for_each_vma(vmi, vma)
599 vsize += vma->vm_end - vma->vm_start;
600 mmap_read_unlock(mm);
601 }
602
603 spin_lock_irq(¤t->sighand->siglock);
604 if (group_dead)
605 pacct->ac_mem = vsize / 1024;
606 if (thread_group_leader(current)) {
607 pacct->ac_exitcode = exitcode;
608 if (current->flags & PF_FORKNOEXEC)
609 pacct->ac_flag |= AFORK;
610 }
611 if (current->flags & PF_SUPERPRIV)
612 pacct->ac_flag |= ASU;
613 if (current->flags & PF_DUMPCORE)
614 pacct->ac_flag |= ACORE;
615 if (current->flags & PF_SIGNALED)
616 pacct->ac_flag |= AXSIG;
617
618 task_cputime(current, &utime, &stime);
619 pacct->ac_utime += utime;
620 pacct->ac_stime += stime;
621 pacct->ac_minflt += current->min_flt;
622 pacct->ac_majflt += current->maj_flt;
623 spin_unlock_irq(¤t->sighand->siglock);
624 }
625
slow_acct_process(struct pid_namespace * ns)626 static void slow_acct_process(struct pid_namespace *ns)
627 {
628 for ( ; ns; ns = ns->parent) {
629 struct bsd_acct_struct *acct = acct_get(ns);
630 if (acct) {
631 do_acct_process(acct);
632 mutex_unlock(&acct->lock);
633 acct_put(acct);
634 }
635 }
636 }
637
638 /**
639 * acct_process - handles process accounting for an exiting task
640 */
acct_process(void)641 void acct_process(void)
642 {
643 struct pid_namespace *ns;
644
645 /*
646 * This loop is safe lockless, since current is still
647 * alive and holds its namespace, which in turn holds
648 * its parent.
649 */
650 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
651 if (ns->bacct)
652 break;
653 }
654 if (unlikely(ns))
655 slow_acct_process(ns);
656 }
657